Object Detection with DeepLabV3 on CityScapes

This repository is for homework 4 of the course Computer Vision. Written by Zhengyuan Su.

Data Preparation

csDownload -d ./data gtFine_trainvaltest.zip
csDownload -d ./data leftImg8bit_trainvaltest.zip

cd data
unzip gtFine_trainvaltest.zip
unzip leftImg8bit_trainvaltest.zip

export CITYSCAPES_DATASET=$(realpath .)
csCreateTrainIdLabelImgs

Model Training and Evaluation

python main.py --tag DeepLabv3 --gpus 0,1,2,3,4,5 # use cross-validation on training set
python main.py --tag DeepLabv3 --gpus 0,1,2,3,4,5 --test # train a whole model, report metrics on the validation set (used as the test set)

I used 6 RTX 3090 to train and one model takes about 1 hour. (Hence to run cross-validation with 5 splits takes 5 hours or so. )

After running, the logs will be synchronized online, and the checkpoints can be found under logs/$TAG/DeepLabv3+/$LOGINDEX/checkpoints.

To evaluate a model, run

python main.py --eval --gpus 7 --eval_ckpt $PATH_TO_CHECKPOINT

Visualization

To visualize, run

python main.py --vis --num_vis 10 --gpus 7 --eval_ckpt $PATH_TO_CHECKPOINT

The result will be saved in ./visualizations.